# Review
> Review a code diff when asked, before committing, or before a PR. Uses scoped evidence, risk lanes, bounded correction, and cold-context escalation. NOT browser QA, prepared-plan critique, or external artifact trust checks.
- Skill: `panda850819/review` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add panda850819/review`
- Raw SKILL.md: https://api.skillmd.com/api/skills/panda850819/review/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: panda850819 (https://skillmd.com/u/panda850819)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/panda850819/review
---
# Code Review
The delta beyond an unstructured single-pass review is scope provenance,
risk-adaptive lanes, evidence gates, bounded correction, and cold escalation.
## 1. Bind scope
1. Read `AGENTS.md` or `CLAUDE.md` when present.
2. Resolve the base from the PR, upstream, or merge-base; print the base and
changed-file list. Never silently review the whole repository.
3. Read the issue, brief, or request that defines intent. If none exists, report
`INTENT GAP` and infer only from the diff.
4. Check branch state and uncommitted changes; use history only to explain scope.
## 2. Choose the risk lane
- **low:** local, reversible, with no trust boundary, persistent data,
concurrency, public interface, generated contract, or production operation.
- **medium:** shared behavior, multiple files, compatibility, or concurrency.
- **high:** auth, secrets, money, permissions, migrations, destructive writes,
production infrastructure, or unfamiliar costly behavior.
Choose the lane from the diff and intent. Promote when the first pass reveals a
risk trigger, candidate finding, coverage gap, or scope drift.
### Low-risk fast path
Only in the low lane with explicit scope and intent: trace the changed path,
match acceptance and branches to tests, and self-refute the likeliest defect. An
`INTENT GAP` promotes rather than entering this fast path. If no finding, gap, or
drift remains: Do not load review learnings; enumerate no lenses,
print no empty scope-drift or cold-review fields, and return:
```markdown
Review scope: .. | files | risk: low
No actionable findings.
Coverage:
Self-refute:
```
## 3. Escalated review
Read `lib/learning-recall.md` and apply relevant repo learnings. Map the diff to
security, data integrity, concurrency, architecture, or operations lenses.
Medium uses only triggered lenses. High uses every relevant lens plus a
cold review.
Trace changed inputs through code, callers, contracts, tests, and failure
handling. A finding survives only with severity `P0`–`P3`, a tight file/line
range, a trigger, mechanism and user-visible consequence, and a correction
direction. Exclude style,
unreachable speculation, and pre-existing defects; review does not edit code.
Use an isolated read-only reviewer when the lane is high, the diff exceeds
roughly 5K tokens, or a load-bearing conclusion remains disputed. Give it only
the bound diff, intent, and applicable repository contract, without the author's
analysis or proposed verdict. Any fresh context that satisfies those semantics is
valid; the transport and model family do not define the review. If no isolated
context is available, record `Cold review: unavailable — no isolation capability`.
A second pass in this same context is not a cold review. Merge findings by
mechanism; disagreement becomes `NEEDS TRACE`, not a vote.
Match acceptance and branches to tests, run the narrowest available checks, and
self-refute the highest-risk assumption. Report `COVERAGE GAP` only for unproved
concrete behavior and `SCOPE DRIFT` only for changes outside intent. Conclude with
findings, `No actionable findings.`, or `BLOCKED`.
## Output and completion
```markdown
Review scope: .. | files | risk:
Findings
- [P1] — <file:line>
Trigger: <input/state>
Mechanism: <why it fails and impact>
Direction: <correction>
Coverage: <verified checks or concrete gaps>
Scope drift: <none or entries>
Cold review: <not earned | completed | unavailable>
Self-refute: <assumption and result>
```
Done when findings are evidence-backed and the report names applicable
coverage, scope drift, cold review, and self-refutation.
## Anti-patterns
- Fixed three-pass fan-out for a small reversible diff.
- Skipping scope provenance or risk promotion because the diff looks small.
- Treating a repository-wide scan as evidence about the changed path.
- Auto-fixing findings and reviewing the rewrite as independent proof.
- Reporting hypothetical security language without an attacker-controlled path.
- Calling an unavailable cold review clean; record the gap.